1:00pm - 1:15pm
Energy demand prediction with machine learning supported by auto-tuning: a case study
1The University of Tokyo, Japan; 2Tokyo University of Science, Japan
The operational energy of buildings is making up one of the highest proportions of life-cycle carbon emissions. A more efficient operation of facilities would result in significant energy savings but necessitates computational models to predict a building’s future energy demands with high precision. To this end, various machine learning models have been proposed in recent years. These models’ prediction accuracies, however, strongly depend on their internal structure and hyperparameters. The time consumption and expertise required for their finetuning call for a more efficient solution. In the context of a case study, this paper describes the relationship between a machine learning model’s prediction accuracy and its hyperparameters. Based on time-stamped recordings of outdoor temperatures and energy demands of a hospital in Japan, recorded every 30 minutes for more than four years, using a deep neural network (DNN) ensemble model, electricity demands were predicted for sixty time steps to follow. Using different evaluation metrics, the effect of hyperparameters on the resulting prediction accuracies was examined. Specifically, automatic hyperparameter tuning methods, such as grid search, random search, and Bayesian optimization, were compared to non-optimized tuning. The results attest to machine learning models’ reliance on hyperparameters and the effectiveness of their automatic tuning.
1:15pm - 1:30pm
Anomaly detection and missing data imputation in building energy data for automated data pre-processing
1Department of Architecture, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan; 2Institute of Industrial Science, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan; 3Tokyo University of Science, 6-3-1, Niijuku, Katsushika-ku, Tokyo 125-8585, Japan
A new trend in building automation is the implementation of smart energy management systems to measure and control building systems without a need for decision-making by human operators. Artificial intelligence can optimize these systems by predicting future demand to make informed decisions about how to efficiently operate individual equipment. These machine learning algorithms use historical data to learn demand trends and require high quality datasets in order to make accurate predictions. But because of issues with data transmission or sensor errors, real world datasets often contain outliers or have data missing. In most research settings, these values can be simply omitted, but in practice, anomalies compromise the automation system’s prediction accuracy, rendering it unable to maximize energy savings. This study explores different machine learning algorithms for anomaly detection for automatically pre-processing incoming data using a case study on an actual electrical demand in a hospital building in Japan, namely cluster-based techniques such as k-means clustering and neural network-based approaches such as the autoencoder. Once anomalies were identified, the missing data was filled with prediction values from a deep neural network model. The newly composed data was then evaluated based on detection accuracy, prediction accuracy and training time. The proposed method of processing anomaly values allows the prediction model to process collected data without interruption, and shows similar predictive accuracy as manually processing the data. These predictions allow energy systems to optimize HVAC equipment control, increasing energy savings and reducing peak building loads.
1:30pm - 1:45pm
Designing a 24 h perturbation method for highly accurate estimation of a building heat transfer coefficient
Université SavoieMont-Blanc, France
In a world-wide effort to decrease the carbon footprint of the existing building stock, massive and extensive building retrofit operations are to be expected. Performance contracting could benefit the investors by securing an effective thermal performance. Verification of the actual thermal performance of the envelope after renovation from on site measurements such as the estimation of the overall Heat Transfer Coefficient is then necessary. Reliable and accurate existing methods to estimate the Heat Transfer Coefficient however need several days to several weeks of undisturbed measurements which can be rather inconvenient for building occupants and quite expensive in terms of operational costs.
This paper intends to design a 24 h heat input signal that ensures the accuracy and precision of estimations of the heat transfer coefficient. A literature review is first provided on existing perturbation methods. On this basis, pseudo-random based heating signals were found promising to achieve accuracy in a short measurement duration. The paper then proposes a numerical methodology to optimize the design of a 24 h signal, inspired by pseudo-random heating signals. The numerical methodology is applied to design a heating signal for a highly insulated wooden frame house.
From preliminary results can be inferred that pseudo-random heating signals achieving accurate estimations in highly insulated wooden frame buildings all share common characteristics : longest heating duration and longest free-floating duration must respectively exceed 7 h and 3 h. The paper finally presents and discusses the results of an experimental validation using one of the well-performing signals.